Camera Models#

COLMAP implements different camera models of varying complexity. If no intrinsic parameters are known a priori, it is generally best to use the simplest camera model that is complex enough to model the distortion effects:

  • SIMPLE_PINHOLE, PINHOLE: Use these camera models, if your images are undistorted a priori. These use one and two focal length parameters, respectively. Note that even in the case of undistorted images, COLMAP could try to improve the intrinsics with a more complex camera model.

  • SIMPLE_RADIAL, RADIAL: This should be the camera model of choice, if the intrinsics are unknown and every image has a different camera calibration, e.g., in the case of Internet photos. Both models are simplified versions of the OPENCV model only modeling radial distortion effects with one and two parameters, respectively.

  • OPENCV, FULL_OPENCV: Use these camera models, if you know the calibration parameters a priori. You can also try to let COLMAP estimate the parameters, if you share the intrinsics for multiple images. Note that the automatic estimation of parameters will most likely fail, if every image has a separate set of intrinsic parameters.

  • SIMPLE_RADIAL_FISHEYE, RADIAL_FISHEYE, OPENCV_FISHEYE, FOV, THIN_PRISM_FISHEYE, RAD_TAN_THIN_PRISM_FISHEYE: Use these camera models for fisheye lenses and note that all other models are not really capable of modeling the distortion effects of fisheye lenses. The FOV model is used by Google Project Tango (make sure to not initialize omega to zero).

  • SIMPLE_FISHEYE, FISHEYE: Use these camera models for fisheye lenses with equidistant projection where distortion can be ignored or has been pre-corrected. These models use the equidistant projection (theta = atan(r)) without any distortion parameters. SIMPLE_FISHEYE has a single focal length (f), while FISHEYE has two (fx, fy).

  • SIMPLE_DIVISION, DIVISION: Use these camera models, if you know the calibration parameters a priori. Similar to SIMPLE_RADIAL and RADIAL models, they can model simple radial distortion effects. The two models have first-order local equivalence for small distortions.

  • EUCM: Use this camera model for wide-angle fisheye cameras and catadioptric systems. It represents radial distortion using two parameters in addition to the standard pinhole parameters.

You can inspect the estimated intrinsic parameters by double-clicking specific images in the model viewer or by exporting the model and opening the cameras.txt file.

Projection#

All perspective camera models map a 3D point in the camera coordinate system to a 2D pixel coordinate in three steps: perspective division, distortion, and the intrinsic transform (focal length and principal point). COLMAP uses a corner-based pixel convention, in which the center of the top-left pixel is at (0.5, 0.5) (see Database Format).

Taking SIMPLE_RADIAL (parameter list f, cx, cy, k) as a worked example, a point \((X, Y, Z)\) in the camera frame, which looks down the positive \(Z\) axis, is projected as follows:

  1. Perspective division onto the normalized image plane:

    \[u = X / Z, \qquad v = Y / Z\]

  2. Radial distortion with \(r^2 = u^2 + v^2\):

    \[u' = u \, (1 + k \, r^2), \qquad v' = v \, (1 + k \, r^2)\]

  3. Focal length and principal point, giving the pixel coordinate:

    \[x = f \, u' + c_x, \qquad y = f \, v' + c_y\]

The inverse mapping (pixel to normalized camera ray) subtracts the principal point, divides by the focal length, and then removes the distortion iteratively.

All other perspective models share this three-step structure and differ only in the number of focal length parameters (a single shared f or separate fx, fy) and in the distortion function, e.g. RADIAL adds a second radial term k2 and OPENCV adds tangential terms p1, p2. The fisheye models instead replace the perspective division with an equidistant projection. The exact parameter list of every model is given by its params_info string and defined in the camera models header: colmap/colmap

Configuration#

To achieve optimal reconstruction results, you might have to try different camera models for your problem. Generally, when the reconstruction fails and the estimated focal length values / distortion coefficients are grossly wrong, it is a sign of using a too complex camera model. Contrary, if COLMAP uses many iterative local and global bundle adjustments, it is a sign of using a too simple camera model that is not able to fully model the distortion effects.

You can also share intrinsics between multiple images to obtain more reliable results (see Share intrinsic camera parameters) or you can fix the intrinsic parameters during the reconstruction (see Fix intrinsic camera parameters).